Model performance can degrade over time as real-world data: Why does responsible AI governance call for

Deployed models drift as data changes, so continuous monitoring and scheduled retraining preserve accuracy and reliability.

The question

Preparing for AIGP? Take the free 5-min readiness quiz →

1. Once validated at launch, a model's behavior remains fixed, so monitoring is only needed if end users happen to file formal complaints about it.

   Wrong because model performance is not fixed; it can drift even without any user complaints.
2. Model performance can degrade over time as real-world data drifts, so ongoing monitoring and scheduled retraining maintain accuracy and reliability. ✓

   Correct because data drift and changing conditions erode performance, which monitoring and retraining counteract.
3. Monitoring is primarily a marketing exercise used to demonstrate engagement metrics and adoption rates to executive and investor audiences over time.

   Wrong because monitoring is a risk and quality control, not a marketing metrics function.
4. Retraining should occur only when the underlying software framework releases a new version, regardless of how the model performs in live production.

   Plausible but wrong because retraining is driven by model performance and drift, not framework release cycles.

How to remember it

Deployed models drift as data changes, so continuous monitoring and scheduled retraining preserve accuracy and reliability.

More Understanding How to Govern AI Deployment and Use questions

• Retrieval augmented generation: Which deployment technique augments a model's responses with information →
• A documented deactivation procedure with clear authority: Which control best enables this? →
• Rights to necessary technical documentation: Which combination of terms best protects it? →
• All 135 Understanding How to Govern AI Deployment and Use questions →